Weekly Run Review

Find the strongest completed runs cheaply, then spend model review only on the five deterministic finalists. Evidence collection is read-only. The final publication owns only one two-line block in AGENTS.md; the full weekly note stays in the local, ignored notes/best_agent_week_<ISO-year>-W<week>.md.

Boundaries

• Never edit campaigns/, domains/, controller evidence, claims, promotions or per-run WRI archives.
• Never launch a product run, choose an experiment for the campaign owner, or change a pass.
• The M/W/F full-run audit owns missing WRI analysis. This weekly flow records a missing synthesis; it does not launch WRI to fill the gap.
• Luna reports are research. Reconcile material numbers against the selection JSON and the cited recorded evidence before publishing them.
• Never expose verifier issue/remedy text, counterexamples, reference artefacts, raw traces, task IDs, per-task failure locations or model reasoning.

references/trajectory.md defines the week-over-week trajectory contract.

1. Run the selector

Start from a fresh worktree at the published main revision. The selector reads the campaign tree of the --repo checkout's main checkout, and the WRI archives in the --repo checkout's local notes/runs/.

bun .claude/skills/weekly-run-review/scripts/select-best-runs.ts \
  --repo /absolute/main/worktree \
  --week previous --timezone Europe/Oslo \
  --top 5 --min-duration-minutes 30 \
  --out /absolute/scratch/weekly-selection.json \
  --luna-manifest /absolute/scratch/luna-sessions.json \
  > /absolute/scratch/weekly-selection.md

--out records the selection as JSON (weekly-best-run-selection/v2), and the console prints the Markdown shortlist, or the same JSON under --json. --luna-manifest writes the launcher manifest and puts the shared instructions beside it as luna-sessions.instructions.md. --now <ISO> pins the census to an instant, and a terminal written after that instant is refused as terminal-in-future, so a pinned census stays that instant's census.

The week is the previous Oslo Monday 00:00 to Monday 00:00, represented as a half-open UTC interval. Completed-run membership uses the controller terminal's writtenAt; wall duration is terminal minus opening. A run exactly at the minimum duration is admitted. Shorter runs remain in the census but consume no WRI or Luna work, and unfinished runs are listed by their opening.

Every run is read through the controller's own strict evidence reader. A run that reader refuses — a terminal that does not bind its opening, a lock without its token, a missing battery record — is listed with the controller's refusal and never re-read leniently, because a lenient reader would rank counts the controller will not stand behind. The seat facts of each duration-eligible run come in process from the outcome readers of the checkout the selector runs in: the metrics bun run outcome prints, and the scorecard's learning yield. A run whose controller recorded no case denominator is a denominator-unrecorded blocker rather than a row ranked at zero. Do not reproduce these readers in the weekly skill.

The selector fills five seats from deterministic evidence before looking for a synthesis. The axes are candidate promotion, vertical completion, informative difficulty, operational yield and non-saturated scale. Exact wins and losses are adjudicated later from the published synthesis; a promotion remains a promotion signal. This is an explicit review queue rather than a hidden weighted capability score. Duration is only an admission filter. Perfect and all-fail batteries are saturated; non-results never enter capability rates.

2. Bind the five main syntheses

For exactly the five selected rows, resolve notes/runs/<runName>/review.json and main_synthesis.md from that local notes/runs/. Bind only a wri-archive/v2 packet whose run ID and full 40-character source revision match the run; any other packet is synthesis-missing. Hash the synthesis bytes. Do not silently replace a missing finalist with rank six.

If a selected synthesis is missing, conflicting or source-mismatched, report the typed gap and stop its Luna admission. The separate full-run audit supplies the missing analysis. Never read a candidate worktree's notes as a substitute for main.

3. Investigate the top five with Luna xhigh

Use codex-luna-swarm's direct launcher because xhigh is explicit:

bun .agents/skills/codex-luna-swarm/scripts/luna-sessions.ts \
  --manifest /absolute/scratch/luna-sessions.json \
  --reasoning-effort xhigh \
  --max-active 4 --start-interval-ms 2000

For five ready finalists the selector generates sixteen distinct sessions: two per run (mechanism and counterfactual) and six cross-run challenges (comparability, full vertical, difficulty, attributable improvement, decision yield, and blind best-run/action review). Do not pad the count. Do not launch the manifest unless the selection's luna.launchAllowed is true; a blocked manifest contains no sessions. Drain completed reports with the launcher's --drain mode while it runs, then once after exit. The launcher gives each session one further attempt itself. After a typed 429, rerun the missing sessions once with the launcher's --retry <outputDir> at lower concurrency and slower pacing. Treat any remaining transport failure as missing research.

Each leaf is read-only and independent. It must state a rival explanation and falsifier, keep exact denominators, and return unobservable rather than guess.

The Scheduled task's primary synthesis runs as gpt-6.1-sol at xhigh. It owns the week-over-week synthesis after the Luna reports return. When this Skill is invoked under another condition, use one bounded read-only Sol xhigh session for that synthesis. Do not ask a Luna session to confirm its own WRI conclusion. If missing finalist syntheses block Luna, the Sol primary still publishes the deterministic trajectory, with semantic mechanisms marked unobservable; it does not fill the missing WRI review.

4. Decide what “best” means

Adjudicate the Luna challenges against deterministic facts and attribution-and-proof. Do not pool different task sets, source identities, model conditions, verifier rules or provenance conditions.

A unique best overall needs coherent identity, a recorded denominator, measurement that supports a decision, claimable evidence and no unresolved product-side blocker. Otherwise publish no claimable winner and name separate axis leaders. A perfect battery means no limit was found. A promotion whose claim is refused remains promoted and unclaimed.

For the chosen run, produce a learning handoff with four things:

1. the mechanisms the next condition should preserve;
2. the most important limitation to target;
3. one bounded counterfactual condition that changes only that limitation; and
4. the evidence that would falsify the proposed explanation.

The chosen run becomes a reference condition rather than an automatic template or launch. The campaign owner decides whether to register the next experiment.

Hand useful tool lessons downstream

Add one delimited Builder-tool automation hand-off section to the weekly note. It is a small hypothesis queue for the separate review-and-improve-builder-tools Automation, not evidence or source-change authority. Include at most three items which concern a Builder-owned tool interface and either recur across two finalists or directly affect one finalist's promotion or completed vertical.

Each item names preserve, investigate or no-action, the exact tool or owner, the observed fact and denominator, safe run/source evidence, current-stack status, one rival explanation and a falsifier. A preserve item protects a successful sequence from an apparently simpler change. An investigate item only prioritises the downstream review; that Automation must prove it from its own source and receipt window before acting. Exclude general model quality, difficulty design, verifier truth and source defects owned outside Builder tools. When no item qualifies, publish No builder-tool hand-off with the missing evidence rather than filling the section speculatively.

5. Publish the weekly reference

Write notes/best_agent_week_<ISO-year>-W<week>.md with the period, full top-five table, winner or no-winner result, mechanisms, limitations, rival explanations, falsifiers, exact identities and denominators, safe evidence paths, Luna result counts, the learning handoff, and this stable block:

<!-- builder-tool-handoff:start -->
## Builder-tool automation hand-off
<zero to three bounded items, or the explicit no-hand-off result>
<!-- builder-tool-handoff:end -->

Always include the delimited weekly-system-trajectory block defined by the trajectory reference. A malformed prior trajectory is an explicit history gap and is never repaired by guessing.

Update only the block delimited by weekly-best-run:start and weekly-best-run:end in AGENTS.md. It has exactly two visible lines:

Best run of <ISO week>: <run, or no claimable winner>.
Why: <at most one concise sentence; the weekly note is local, so it links nothing under notes/>.

If the marker is absent, duplicated or malformed, stop instead of rewriting the file broadly. A rerun may update the same week's note; it never rewrites an earlier week's note. Check git diff --check, review the AGENTS.md diff, commit only that file, refresh origin/main, replay if necessary, and push HEAD:main under the repository's surrounding-file rule. A non-fast-forward or overlapping marker change is a publication conflict and does not authorise an overwrite.

Return the same concise report in the Scheduled task. Link the weekly note, selected main_synthesis.md, digest.md and review.json files; omit raw Luna reports.